Face Generation

In this project, you'll define and train a Generative Adverserial network of your own creation on a dataset of faces. Your goal is to get a generator network to generate new images of faces that look as realistic as possible!

The project will be broken down into a series of tasks from defining new architectures training adversarial networks. At the end of the notebook, you'll be able to visualize the results of your trained Generator to see how it performs; your generated samples should look like fairly realistic faces with small amounts of noise.

Get the Data

You'll be using the CelebFaces Attributes Dataset (CelebA) to train your adversarial networks.

This dataset has higher resolution images than datasets you have previously worked with (like MNIST or SVHN) you've been working with, and so, you should prepare to define deeper networks and train them for a longer time to get good results. It is suggested that you utilize a GPU for training.

Pre-processed Data

Since the project's main focus is on building the GANs, we've done some of the pre-processing for you. Each of the CelebA images has been cropped to remove parts of the image that don't include a face, then resized down to 64x64x3 NumPy images. Some sample data is show below.

If you are working locally, you can download this data by clicking here

This is a zip file that you'll need to extract in the home directory of this notebook for further loading and processing. After extracting the data, you should be left with a directory of data processed-celeba-small/.

# run this once to unzip the file
!unzip processed-celeba-small.zip

from typing import Tuple, Callable

import matplotlib.pyplot as plt
import numpy as np
from torch.utils.data import DataLoader
from torchvision.transforms import (
    Compose,
    ToTensor,
    Normalize,
    Resize
)

import tests

data_dir = 'processed_celeba_small/celeba/'

Data pipeline

The CelebA dataset contains over 200,000 celebrity images with annotations. Since you're going to be generating faces, you won't need the annotations, you'll only need the images. Note that these are color images with 3 color channels (RGB)#RGB_Images) each.

Pre-process and Load the Data

Since the project's main focus is on building the GANs, we've done some of the pre-processing for you. Each of the CelebA images has been cropped to remove parts of the image that don't include a face, then resized down to 64x64x3 NumPy images. This pre-processed dataset is a smaller subset of the very large CelebA dataset and contains roughly 30,000 images.

Your first task consists in building the dataloader. To do so, you need to do the following:

• implement the get_transforms function
• create a custom Dataset class that reads the CelebA data

Exercise: implement the get_transforms function

The get_transforms function should output a torchvision.transforms.Compose of different transformations. You have two constraints:

• the function takes a tuple of size as input and should resize the images to the input size
• the output images should have values ranging from -1 to 1

def get_transforms(size: Tuple[int, int]) -> Callable:
    """
    Transforms to apply to the image.
    Args:
        size: Target size of the images
    Returns:
        Composition of transforms
    """
    transforms = [
        Resize(size),
        ToTensor(),
        Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
    ]
    return Compose(transforms)

Exercise: implement the DatasetDirectory class

The DatasetDirectory class is a torch Dataset that reads from the above data directory. The __getitem__ method should output a transformed tensor and the __len__ method should output the number of files in our dataset. You can look at this custom dataset for ideas.

from PIL import Image
import torch
from torch.utils.data import Dataset
import os
from typing import Callable


class DatasetDirectory(Dataset):
    def __init__(self, directory: str, transforms: Callable = None, extension: str = '.jpg'):
        """Initialize dataset with directory of images."""
        self.directory = directory
        self.image_files = sorted([
            os.path.join(directory, f)
            for f in os.listdir(directory)
            if f.endswith(extension)
        ])

        # Default transform if none provided
        self.transforms = transforms if transforms else get_transforms((64, 64))

    def __len__(self) -> int:
        return len(self.image_files)

    def __getitem__(self, index: int) -> torch.Tensor:
        """Get a single image."""
        # Get image path
        image_path = self.image_files[index]

        # Load and convert image to RGB
        image = Image.open(image_path).convert('RGB')

        # Apply transforms
        if self.transforms:
            image = self.transforms(image)

        return image

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""
dataset = DatasetDirectory(data_dir, get_transforms((64, 64)))
tests.check_dataset_outputs(dataset)

Congrats, your dataset implementation passed all the tests

The functions below will help you visualize images from the dataset.

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""


def denormalize(images):
    """Transform images from [-1.0, 1.0] to [0, 255] and cast them to uint8."""
    return ((images + 1.) / 2. * 255).astype(np.uint8)


# plot the images in the batch, along with the corresponding labels
fig = plt.figure(figsize=(20, 4))
plot_size = 20
for idx in np.arange(plot_size):
    ax = fig.add_subplot(2, int(plot_size / 2), idx + 1, xticks=[], yticks=[])
    img = dataset[idx].numpy()
    img = np.transpose(img, (1, 2, 0))
    img = denormalize(img)
    ax.imshow(img)

Model implementation

As you know, a GAN is comprised of two adversarial networks, a discriminator and a generator. Now that we have a working data pipeline, we need to implement the discriminator and the generator.

Feel free to implement any additional class or function.

Exercise: Create the discriminator

The discriminator's job is to score real and fake images. You have two constraints here:

• the discriminator takes as input a batch of 64x64x3 images
• the output should be a single value (=score)

Feel free to get inspiration from the different architectures we talked about in the course, such as DCGAN, WGAN-GP or DRAGAN.

Some tips

• To scale down from the input image, you can either use Conv2d layers with the correct hyperparameters or Pooling layers.
• If you plan on using gradient penalty, do not use Batch Normalization layers in the discriminator.

from torch.nn import Module
import torch
import torch.nn as nn

class Discriminator(Module):
    def __init__(self):
        super(Discriminator, self).__init__()

        # DCGAN-style discriminator
        self.main = nn.Sequential(
            # Input: 3x64x64
            nn.Conv2d(3, 64, 4, stride=2, padding=1, bias=False),
            nn.LeakyReLU(0.2, inplace=True),

            # 64x32x32
            nn.Conv2d(64, 128, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(128),
            nn.LeakyReLU(0.2, inplace=True),

            # 128x16x16
            nn.Conv2d(128, 256, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(256),
            nn.LeakyReLU(0.2, inplace=True),

            # 256x8x8
            nn.Conv2d(256, 512, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(512),
            nn.LeakyReLU(0.2, inplace=True),

            # 512x4x4 -> 1x1x1x1
            nn.Conv2d(512, 1, 4, stride=1, padding=0, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.main(x).view(-1, 1, 1, 1)

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""
discriminator = Discriminator()
tests.check_discriminator(discriminator)

Congrats, your discriminator implementation passed all the tests

Exercise: create the generator

The generator's job creates the "fake images" and learns the dataset distribution. You have three constraints here:

• the generator takes as input a vector of dimension [batch_size, latent_dimension, 1, 1]
• the generator must outputs 64x64x3 images

Feel free to get inspiration from the different architectures we talked about in the course, such as DCGAN, WGAN-GP or DRAGAN.

Some tips:

• to scale up from the latent vector input, you can use ConvTranspose2d layers
• as often with Gan, Batch Normalization helps with training

class Generator(Module):
    def __init__(self, latent_dim: int):
        super(Generator, self).__init__()

        self.main = nn.Sequential(
            # Input: latent_dim x 1 x 1
            nn.ConvTranspose2d(latent_dim, 512, 4, stride=1, padding=0, bias=False),
            nn.BatchNorm2d(512),
            nn.ReLU(True),

            # 512 x 4 x 4
            nn.ConvTranspose2d(512, 256, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(256),
            nn.ReLU(True),

            # 256 x 8 x 8
            nn.ConvTranspose2d(256, 128, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(128),
            nn.ReLU(True),

            # 128 x 16 x 16
            nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(True),

            # 64 x 32 x 32
            nn.ConvTranspose2d(64, 3, 4, stride=2, padding=1, bias=False),
            nn.Tanh()
            # Output: 3 x 64 x 64
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.main(x)

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""
# run this cell to verify your generator implementation
# latent_dim = 128
latent_dim = 256
generator = Generator(latent_dim)
tests.check_generator(generator, latent_dim)

Congrats, your generator implementation passed all the tests

Optimizer

In the following section, we create the optimizers for the generator and discriminator. You may want to experiment with different optimizers, learning rates and other hyperparameters as they tend to impact the output quality.

Exercise: implement the optimizers

import torch.optim as optim


def create_optimizers(generator: Module, discriminator: Module):
    """Creates optimizers for generator and discriminator"""
    
    g_optimizer = optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))
    d_optimizer = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.999))  # Lowered discriminator LR

    return g_optimizer, d_optimizer

Losses implementation

In this section, we are going to implement the loss function for the generator and the discriminator. You can and should experiment with different loss function.

Some tips:

• You can choose the commonly used the binary cross entropy loss or select other losses we have discovered in the course, such as the Wasserstein distance.
• You may want to implement a gradient penalty function as discussed in the course. It is not required and the code will work whether you implement it or not.

Exercise: implement the generator loss

The generator's goal is to get the discriminator to think its generated images (= "fake" images) are real.

def generator_loss(fake_logits):
    """Generator loss"""
    criterion = nn.BCELoss()
    # Generate labels for fake images as if they were real
    labels = torch.ones_like(fake_logits)
    return criterion(fake_logits, labels)

Exercise: implement the discriminator loss

We want the discriminator to give high scores to real images and low scores to fake ones and the discriminator loss should reflect that.

def discriminator_loss(real_logits, fake_logits):
    """Discriminator loss"""
    criterion = nn.BCELoss()
    # Real images should be labeled as 1
    real_labels = torch.ones_like(real_logits)
    # Fake images should be labeled as 0
    fake_labels = torch.zeros_like(fake_logits)

    real_loss = criterion(real_logits, real_labels)
    fake_loss = criterion(fake_logits, fake_labels)

    return real_loss + fake_loss

Exercise (Optional): Implement the gradient Penalty

In the course, we discussed the importance of gradient penalty in training certain types of Gans. Implementing this function is not required and depends on some of the design decision you made (discriminator architecture, loss functions).

def gradient_penalty(discriminator, real_samples, fake_samples):
    """
    Calculate the gradient penalty loss for WGAN GP
    Args:
        discriminator: the discriminator model
        real_samples: real images from dataset
        fake_samples: generated images from generator
    Returns:
        gradient penalty value
    """
    alpha = torch.rand((real_samples.size(0), 1, 1, 1)).to(real_samples.device)
    
    interpolates = (alpha * real_samples + ((1 - alpha) * fake_samples)).requires_grad_(True)
    
    d_interpolates = discriminator(interpolates)
    
    gradients = torch.autograd.grad(
        outputs=d_interpolates,
        inputs=interpolates,
        grad_outputs=torch.ones_like(d_interpolates).to(real_samples.device),
        create_graph=True,
        retain_graph=True,
        only_inputs=True
    )[0]
    
    gradients = gradients.view(gradients.size(0), -1)
    gradient_norm = gradients.norm(2, dim=1)
    gradient_penalty = ((gradient_norm - 1) ** 2).mean()
    
    return gradient_penalty

Training

Training will involve alternating between training the discriminator and the generator. You'll use your functions real_loss and fake_loss to help you calculate the discriminator losses.

• You should train the discriminator by alternating on real and fake images
• Then the generator, which tries to trick the discriminator and should have an opposing loss function

Exercise: implement the generator step and the discriminator step functions

Each function should do the following:

• calculate the loss
• backpropagate the gradient
• perform one optimizer step

def generator_step(generator, discriminator, g_optimizer, batch_size, latent_dim, device):
    """One training step for generator"""
    g_optimizer.zero_grad()

    # Generate fake images
    noise = torch.randn(batch_size, latent_dim, 1, 1, device=device)
    fake_images = generator(noise)

    # Get discriminator output on fake images
    fake_logits = discriminator(fake_images)

    # Calculate generator loss
    g_loss = generator_loss(fake_logits)

    # Backpropagate and update
    g_loss.backward()
    g_optimizer.step()

    return {'loss': g_loss}


def discriminator_step(generator, discriminator, d_optimizer, batch_size, latent_dim, real_images, device):
    """One training step for discriminator"""
    d_optimizer.zero_grad()

    # Get discriminator output on real images
    real_logits = discriminator(real_images)

    # Generate and get output on fake images
    noise = torch.randn(batch_size, latent_dim, 1, 1, device=device)
    with torch.no_grad():
        fake_images = generator(noise)
    fake_logits = discriminator(fake_images)

    # Calculate discriminator loss
    d_loss = discriminator_loss(real_logits, fake_logits)

    # Backpropagate and update
    d_loss.backward()
    d_optimizer.step()

    return {'loss': d_loss}

Main training loop

You don't have to implement anything here but you can experiment with different hyperparameters.

latent_dim = 256 

# Use for Apple MPS 
if torch.cuda.is_available():
    device = 'cuda'  
elif torch.backends.mps.is_available():
    device = 'mps'  
else:
    device = 'cpu'  

print(f"Using device: {device}")

n_epochs = 50 

# Larger batch sizes stabilize training (reduce if memory is limited)
batch_size = 128  

Using device: mps

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""
print_every = 50

generator = Generator(latent_dim).to(device)
discriminator = Discriminator().to(device)
g_optimizer, d_optimizer = create_optimizers(generator, discriminator)

dataloader = DataLoader(dataset,
                        batch_size=64,
                        shuffle=True,
                        num_workers=0,
                        drop_last=True,
                        pin_memory=False)

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""


def display(fixed_latent_vector: torch.Tensor):
    """ helper function to display images during training """
    fig = plt.figure(figsize=(14, 4))
    plot_size = 16
    for idx in np.arange(plot_size):
        ax = fig.add_subplot(2, int(plot_size / 2), idx + 1, xticks=[], yticks=[])
        img = fixed_latent_vector[idx, ...].detach().cpu().numpy()
        img = np.transpose(img, (1, 2, 0))
        img = denormalize(img)
        ax.imshow(img)
    plt.show()

Exercise: implement the training strategy

You should experiment with different training strategies. For example:

• train the generator more often than the discriminator.
• added noise to the input image
• use label smoothing

Implement with your training strategy below.

from datetime import datetime
from torch.optim.lr_scheduler import StepLR

# Fixed latent vector for visualizing generator's progress
fixed_latent_vector = torch.randn(16, latent_dim, 1, 1).to(device)

losses = []

# Define the learning rate schedulers
d_scheduler = StepLR(d_optimizer, step_size=30, gamma=0.5)  # Reduce LR every 30 epochs
g_scheduler = StepLR(g_optimizer, step_size=30, gamma=0.5)

for epoch in range(n_epochs):
    print(f"Starting Epoch [{epoch + 1}/{n_epochs}]...")  # Start of each epoch

    for batch_i, real_images in enumerate(dataloader):
        real_images = real_images.to(device)

        ####################################
        # Train the discriminator
        d_loss = discriminator_step(
            generator, discriminator, d_optimizer,
            batch_size, latent_dim, real_images, device
        )

        # Train the generator twice for better convergence
        g_loss = generator_step(
            generator, discriminator, g_optimizer,
            batch_size, latent_dim, device
        )
        g_loss = generator_step(
            generator, discriminator, g_optimizer,
            batch_size, latent_dim, device
        )
        ####################################

        # Append discriminator and generator losses for tracking
        d = d_loss['loss'].item()
        g = g_loss['loss'].item()
        losses.append((d, g))

        # Print loss details only at the end of the epoch
        if batch_i == len(dataloader) - 1:
            time = str(datetime.now()).split('.')[0]
            print(
                f'End of Epoch [{epoch + 1}/{n_epochs}]. {time} | Epoch [{epoch + 1}/{n_epochs}] | Batch {batch_i}/{len(dataloader)} | d_loss: {d:.4f} | g_loss: {g:.4f}')

    # Display images during training
    generator.eval()
    generated_images = generator(fixed_latent_vector)
    display(generated_images)
    generator.train()

    # Step the learning rate schedulers after each epoch
    d_scheduler.step()
    g_scheduler.step()

print("Training complete.")

Starting Epoch [1/50]...
End of Epoch [1/50]. 2024-11-01 18:57:54 | Epoch [1/50] | Batch 508/509 | d_loss: 1.4551 | g_loss: 1.0298

Starting Epoch [2/50]...
End of Epoch [2/50]. 2024-11-01 18:59:39 | Epoch [2/50] | Batch 508/509 | d_loss: 1.5948 | g_loss: 0.5879

Starting Epoch [3/50]...
End of Epoch [3/50]. 2024-11-01 19:02:02 | Epoch [3/50] | Batch 508/509 | d_loss: 1.3511 | g_loss: 0.9748

Starting Epoch [4/50]...
End of Epoch [4/50]. 2024-11-01 19:04:35 | Epoch [4/50] | Batch 508/509 | d_loss: 1.3582 | g_loss: 0.8063

Starting Epoch [5/50]...
End of Epoch [5/50]. 2024-11-01 19:06:38 | Epoch [5/50] | Batch 508/509 | d_loss: 1.2900 | g_loss: 0.9433

Starting Epoch [6/50]...
End of Epoch [6/50]. 2024-11-01 19:08:36 | Epoch [6/50] | Batch 508/509 | d_loss: 1.3693 | g_loss: 0.6492

Starting Epoch [7/50]...
End of Epoch [7/50]. 2024-11-01 19:10:33 | Epoch [7/50] | Batch 508/509 | d_loss: 1.2642 | g_loss: 0.9723

Starting Epoch [8/50]...
End of Epoch [8/50]. 2024-11-01 19:12:31 | Epoch [8/50] | Batch 508/509 | d_loss: 1.4230 | g_loss: 1.0284

Starting Epoch [9/50]...
End of Epoch [9/50]. 2024-11-01 19:14:34 | Epoch [9/50] | Batch 508/509 | d_loss: 1.0509 | g_loss: 1.2985

Starting Epoch [10/50]...
End of Epoch [10/50]. 2024-11-01 19:16:37 | Epoch [10/50] | Batch 508/509 | d_loss: 1.1629 | g_loss: 1.1712

Starting Epoch [11/50]...
End of Epoch [11/50]. 2024-11-01 19:18:39 | Epoch [11/50] | Batch 508/509 | d_loss: 1.4589 | g_loss: 0.4053

Starting Epoch [12/50]...
End of Epoch [12/50]. 2024-11-01 19:20:40 | Epoch [12/50] | Batch 508/509 | d_loss: 0.7663 | g_loss: 0.6661

Starting Epoch [13/50]...
End of Epoch [13/50]. 2024-11-01 19:22:42 | Epoch [13/50] | Batch 508/509 | d_loss: 0.7267 | g_loss: 1.0778

Starting Epoch [14/50]...
End of Epoch [14/50]. 2024-11-01 19:24:43 | Epoch [14/50] | Batch 508/509 | d_loss: 0.8844 | g_loss: 0.9200

Starting Epoch [15/50]...
End of Epoch [15/50]. 2024-11-01 19:26:44 | Epoch [15/50] | Batch 508/509 | d_loss: 0.5894 | g_loss: 1.5995

Starting Epoch [16/50]...
End of Epoch [16/50]. 2024-11-01 19:28:46 | Epoch [16/50] | Batch 508/509 | d_loss: 0.7620 | g_loss: 1.4146

Starting Epoch [17/50]...
End of Epoch [17/50]. 2024-11-01 19:30:48 | Epoch [17/50] | Batch 508/509 | d_loss: 0.5030 | g_loss: 2.5242

Starting Epoch [18/50]...
End of Epoch [18/50]. 2024-11-01 19:32:50 | Epoch [18/50] | Batch 508/509 | d_loss: 0.8147 | g_loss: 5.0975

Starting Epoch [19/50]...
End of Epoch [19/50]. 2024-11-01 19:34:50 | Epoch [19/50] | Batch 508/509 | d_loss: 1.0508 | g_loss: 3.4355

Starting Epoch [20/50]...
End of Epoch [20/50]. 2024-11-01 19:36:51 | Epoch [20/50] | Batch 508/509 | d_loss: 0.4976 | g_loss: 1.3290

Starting Epoch [21/50]...
End of Epoch [21/50]. 2024-11-01 19:38:52 | Epoch [21/50] | Batch 508/509 | d_loss: 0.4508 | g_loss: 2.6244

Starting Epoch [22/50]...
End of Epoch [22/50]. 2024-11-01 19:40:55 | Epoch [22/50] | Batch 508/509 | d_loss: 0.1679 | g_loss: 3.3131

Starting Epoch [23/50]...
End of Epoch [23/50]. 2024-11-01 19:43:03 | Epoch [23/50] | Batch 508/509 | d_loss: 0.6330 | g_loss: 4.9456

Starting Epoch [24/50]...
End of Epoch [24/50]. 2024-11-01 19:45:05 | Epoch [24/50] | Batch 508/509 | d_loss: 0.5138 | g_loss: 4.2691

Starting Epoch [25/50]...
End of Epoch [25/50]. 2024-11-01 19:47:07 | Epoch [25/50] | Batch 508/509 | d_loss: 1.7220 | g_loss: 6.4401

Starting Epoch [26/50]...
End of Epoch [26/50]. 2024-11-01 19:49:09 | Epoch [26/50] | Batch 508/509 | d_loss: 0.8203 | g_loss: 0.9184

Starting Epoch [27/50]...
End of Epoch [27/50]. 2024-11-01 19:51:15 | Epoch [27/50] | Batch 508/509 | d_loss: 0.0818 | g_loss: 3.8678

Starting Epoch [28/50]...
End of Epoch [28/50]. 2024-11-01 19:53:22 | Epoch [28/50] | Batch 508/509 | d_loss: 1.4045 | g_loss: 5.0645

Starting Epoch [29/50]...
End of Epoch [29/50]. 2024-11-01 19:55:25 | Epoch [29/50] | Batch 508/509 | d_loss: 0.3280 | g_loss: 3.7874

Starting Epoch [30/50]...
End of Epoch [30/50]. 2024-11-01 19:57:26 | Epoch [30/50] | Batch 508/509 | d_loss: 0.9399 | g_loss: 3.7612

Starting Epoch [31/50]...
End of Epoch [31/50]. 2024-11-01 19:59:28 | Epoch [31/50] | Batch 508/509 | d_loss: 0.1053 | g_loss: 4.5860

Starting Epoch [32/50]...
End of Epoch [32/50]. 2024-11-01 20:01:32 | Epoch [32/50] | Batch 508/509 | d_loss: 0.0920 | g_loss: 4.4004

Starting Epoch [33/50]...
End of Epoch [33/50]. 2024-11-01 20:03:35 | Epoch [33/50] | Batch 508/509 | d_loss: 0.1588 | g_loss: 3.4051

Starting Epoch [34/50]...
End of Epoch [34/50]. 2024-11-01 20:05:40 | Epoch [34/50] | Batch 508/509 | d_loss: 0.0980 | g_loss: 3.7069

Starting Epoch [35/50]...
End of Epoch [35/50]. 2024-11-01 20:07:46 | Epoch [35/50] | Batch 508/509 | d_loss: 0.1126 | g_loss: 3.9702

Starting Epoch [36/50]...
End of Epoch [36/50]. 2024-11-01 20:09:52 | Epoch [36/50] | Batch 508/509 | d_loss: 0.0961 | g_loss: 5.3729

Starting Epoch [37/50]...
End of Epoch [37/50]. 2024-11-01 20:11:56 | Epoch [37/50] | Batch 508/509 | d_loss: 0.0801 | g_loss: 4.1903

Starting Epoch [38/50]...
End of Epoch [38/50]. 2024-11-01 20:13:58 | Epoch [38/50] | Batch 508/509 | d_loss: 0.0468 | g_loss: 4.8736

Starting Epoch [39/50]...
End of Epoch [39/50]. 2024-11-01 20:16:00 | Epoch [39/50] | Batch 508/509 | d_loss: 0.0459 | g_loss: 5.1289

Starting Epoch [40/50]...
End of Epoch [40/50]. 2024-11-01 20:18:07 | Epoch [40/50] | Batch 508/509 | d_loss: 0.0901 | g_loss: 3.7732

Starting Epoch [41/50]...
End of Epoch [41/50]. 2024-11-01 20:20:13 | Epoch [41/50] | Batch 508/509 | d_loss: 0.0177 | g_loss: 5.1816

Starting Epoch [42/50]...
End of Epoch [42/50]. 2024-11-01 20:22:14 | Epoch [42/50] | Batch 508/509 | d_loss: 0.0361 | g_loss: 5.2459

Starting Epoch [43/50]...
End of Epoch [43/50]. 2024-11-01 20:24:15 | Epoch [43/50] | Batch 508/509 | d_loss: 0.0453 | g_loss: 5.7086

Starting Epoch [44/50]...
End of Epoch [44/50]. 2024-11-01 20:26:16 | Epoch [44/50] | Batch 508/509 | d_loss: 0.0492 | g_loss: 4.4198

Starting Epoch [45/50]...
End of Epoch [45/50]. 2024-11-01 20:28:17 | Epoch [45/50] | Batch 508/509 | d_loss: 0.0769 | g_loss: 3.9152

Starting Epoch [46/50]...
End of Epoch [46/50]. 2024-11-01 20:30:18 | Epoch [46/50] | Batch 508/509 | d_loss: 0.0572 | g_loss: 3.3547

Starting Epoch [47/50]...
End of Epoch [47/50]. 2024-11-01 20:32:19 | Epoch [47/50] | Batch 508/509 | d_loss: 0.0735 | g_loss: 4.4590

Starting Epoch [48/50]...
End of Epoch [48/50]. 2024-11-01 20:34:19 | Epoch [48/50] | Batch 508/509 | d_loss: 0.1185 | g_loss: 6.1540

Starting Epoch [49/50]...
End of Epoch [49/50]. 2024-11-01 20:36:18 | Epoch [49/50] | Batch 508/509 | d_loss: 0.0331 | g_loss: 5.6865

Starting Epoch [50/50]...
End of Epoch [50/50]. 2024-11-01 20:38:18 | Epoch [50/50] | Batch 508/509 | d_loss: 0.0614 | g_loss: 5.7153

Training complete.

Training losses

Plot the training losses for the generator and discriminator.

"""
DO NOT MODIFY ANYTHING IN THIS CELL
"""
fig, ax = plt.subplots(figsize=(16, 4))
losses = np.array(losses)
plt.plot(losses.T[0], label='Discriminator', alpha=0.5)
plt.plot(losses.T[1], label='Generator', alpha=0.5)
plt.title("Training Losses")
plt.legend()

<matplotlib.legend.Legend at 0x387076b20>

!jupyter nbconvert --to html dlnd_face_generation_starter.ipynb

Question: What do you notice about your generated samples and how might you improve this model?

When you answer this question, consider the following factors:

• The dataset is biased; it is made of "celebrity" faces that are mostly white
• Model size; larger models have the opportunity to learn more features in a data feature space
• Optimization strategy; optimizers and number of epochs affect your final result
• Loss functions

Answer: The generated faces show reasonable structure but lack fine details and diversity, with noticeable artifacts and blurriness. The training graphs indicate instability in generator performance (high fluctuating loss) while the discriminator remains relatively stable. The model could be improved by increasing model capacity, implementing WGAN-GP loss with gradient penalty, using progressive growing architecture, and most importantly, training for more epochs with a more diverse dataset to address the current bias towards white celebrities.

Submitting This Project

When submitting this project, make sure to run all the cells before saving the notebook. Save the notebook file as "dlnd_face_generation.ipynb".

Submit the notebook using the SUBMIT button in the bottom right corner of the Project Workspace.